IETF RMCAT WG: Video Quality Metrics Discussion for Evalua?on Criteria V. S. Somayazulu (v.srinivasa.somayazulu@intel.com) Hassnaa Moustafa (hassnaa.moustafa@intel.com)
Problem Statement Video Quality Percep1on is cri?cal for end- user QoE Topic has been discussed in WG, but no conclusion yet on including this in the requirements or the evalua?on metrics for the conges?on control algorithms Network/transport related metrics are currently present in the evalua?on criteria for Rmcat conges?on control algorithms, but do not fully capture the impact on video QoE. Purpose of this presenta/on: 1. Provide background on video quality metrics that can help quan/fy impact of conges/on control on the video QoE 2. S/mulate a discussion on adop/on of appropriate metrics in evalua/on criteria
(High Level) Summary of Evalua?on Scenario Discussions in RMCAT WG Current (evolving) direc?on seems to be - Evaluate overall performance, i.e., Video/Audio+ Network Conges?on control algorithm Couple of ways to include video characteris?cs: 1. Use a traffic model that captures sta?s?cs of the video source+rate control+shaping 2. Possibly use uncompressed video sequences with a live encoder (+ prescribed sezngs) and conges?on control algo. With either approach, we should define a means of capturing impact on the overall video QoE.
Objec?ve Metrics Big Picture View Background Objec?ve metrics developed to mimic human percep?on of video, e.g., look at the received video, and judge how good it looks Distor?ons between source and video considered a black box and did not affect metric design Good for unbiased evalua?on of encoders, etc. Non- reference / Reduced reference a difficult problem in this generic case Full- reference vs. non- reference VQM Full reference: compare the measured video with the original uncompressed video PSNR, MS- SSIM, PEVQ Non- reference: analyze the video without a comparison E.g. P1202, etc. ITU- T G.1070 H.264 Encode Source Video IP (UDP) Packets Packet Loss/Jiier Distor?ons Viewed Video Buffer Management Pkt Loss Concealment H.264 Decode Intel Confiden?al 4
Video Conferencing QoE ITU- T G.1070 ITU- T G.1070 Recommenda?on ITU- T G.1070 proposes an algorithm that es?mates videophone quality for quality of experience (QoE)/quality of service (QoS) planners.
ITU- T G.1070 Video Quality Metric Video quality is calculated as: where I coding represents basic video quality as a func?on of video bitrate and frame rate D PplV represents degree of video quality robustness due to packet loss and PplV is the packet loss rate in % These quan??es are calculated using a set of fixed parameters dependent on codec type, video format, key frame interval, and video display size G.1070 provides provisional values for H.264, VGA format, 1 second key frame interval and 9.2 inch display, coded bit rates between 400 kbps 2 Mbps, packet loss rates < 5% and frame rates from 5-25 fps. Parameter values for modeling any other set of condi?ons would need to be derived from video quality evalua?on Further enhancements to the basic model for different codecs, display formats, content dependency, etc. have been explored, e.g.: [Joskowicz, 2009] Joskowicz, J. et al Enhancements to the Opinion Model for Video- Telephony Applica?ons, Proc. 5 th La5n American Networking Conference, pp. 87-94.
Considera?ons on use of ITU- T G.1070 Use evalua?on scenario(s) to run simula?on of target conges?on avoidance algorithm with a given test video sequence (trace file) Collect trace of packet arrival?mes, packet losses, at sender and receiver Segment data into short intervals of?me (e.g. 5 seconds?) For each segment i: Calculate video bitrate sent, assume fixed frame rate Calculate average packet loss rate at receiver Calculate per- segment video quality V q, i Calculate over the en?re video sequence: Mean and variance of the set of {V q, i } Higher mean and lower variance à beier overall video quality
Discussion Propose to add video quality evalua?on metrics for considera?on by Rmcat WG. Desirable features of the metrics should include: Good correla?on with subjec?ve video quality percep?on Combine different parameters to provide an integrated look at video QoE impact Rela?vely simple to calculate based on data from network simula?ons Ideally, based on published standards Subjec?ve quality evalua?on is hard to organize and execute, especially in a contribu?on evalua?on phase Common objec?ve evalua?on metrics of video quality Easier to use in proposal evalua?ons, Full- reference & Non- reference: laier may be more suited for RMCAT evalua?on ITU- T Rec.G.1070 based NR VQ metrics designed for video conferencing applica?ons Addi?onally, this doesn t require compressed bitstream inspec?on, etc.: algorithm inputs are high level, e.g. throughput, packet loss rate, etc. Solici/ng feedback from group on defining considera/on for video quality metrics as part of evalua/on criteria for conges/on control algorithms? à Current phase For further considera/on: Could VQ be exploited by conges/on control algorithms? Poten/al for incorpora/ng video quality informa/on into RTCP XRBLOCK reports? à (Later phase)
Annex
Background Rmcat WG is dealing with conges?on control for Internet data considering interac?ve point- to- point real-?me mul?media services over RTP Requirements for conges?on control algorithms are defined considering Low delay Semi- Reliable data delivery Fairness to other flows Adapta?on to network condi?ons Metrics for conges?on control are defined to be Delay, throughput, minimizing transmission rates oscilla?ons, reac?vity to transient events and packet losses and discards Evalua?on Criteria for conges?on control algorithms have been defined considering Avoiding Conges?on Collapse Stability Media Traffic Startup Behavior Diverse Environments Varying Path Characteris?cs Reac?ng to Transient Events or Interrup?ons Fairness with Similar Cross Traffic Impact on Cross Traffic Extensions to RTP/RTCP
Video Quality Varia?on w/ Network Condi?ons Varia?on PSNRloss for different network bandwidth limita5ons PSNRloss for different packet loss rate Source: R- G Cheng et al., Measurement and Analysis of Skype Video Traffic, APWCS 2008.
Impact of Throughput Varia?on on Video Conferencing Applica?ons Source: L. De Cicco et al., Skype Video Responsiveness to Bandwidth Varia5on, ACM NOSSDAV 2008.
Video Quality Evalua?on: Introduc?on Quality of Experience The overall experience the consumer has when accessing and using provided video services Quan?fying Video Quality Mean Opinion Score (MOS) Subjec?vely done: recruit a group of people to watch a set of video clips and give a numeric score to each clip Automa?cally done: design algorithms to es?mate a MOS based on characteris?cs of media stream, network, device, etc. Video Quality Issues Video crea?on, video encoding/transcoding, video transmission, video display
Video Quality Issues Video blockiness (encoding) Video blurriness (capturing/encoding) Video losses (transmission) Video jerkiness (transmission/encoding/display) Video freezing/rebuffering (transmission) A/V sync problem (transmission/encoding) Blockiness vs. blurriness
PSNR (Peak Signal- to- Noise Ra?o) Most commonly used metric to measure the quality of reconstruc?on of lossy compression codecs PSNR=10 log 10 ( 255 2 /MSE ) Typically values between 30~50 db, higher is beier
MS- SSIM MSSSIM(x,y)= [ l M (x,y)] α M j=1 M [ c j (x,y)] β j [ s j (x,y)] γ j Luminance l(x, y) Contrast c(x,y) Structure s(x,y) Typical values between 0.8~1, higher is beier
PEVQ Output PEVQ MOS ranging from 1 (bad)~5 (excellent) Distor?on Indicators: Delay, Brightness, Contrast, PSNR, Jerkiness, Blur, Blockiness, Frame skips and freezes, Temporal Ac?vity and spa?al complexity
ITU- T SG12 No- Reference Objec?ve Standards Inputs Consented or Approved documents shown in black Targeted Applica?ons: IPTV services; non- adap?ve streaming; non- interac?ve P.1201.x mul?media QoE; P.1202.x video QoE Targeted protocols: non- HTTP (RTP, TS- on UDP, etc.) Packet Headers P.1201.1 (lbr*) P.1201.2 (hbr*) P.1202.1 (lbr*) P.1202.2 mode 1 (hbr*) P.1202.2 mode 2 (hbr*) Bit Stream J.bitvqm,J.mm-noref, J.noref, VQEG-JEG Hybrid Project Decoded Video * lbr- low bit rate hbr- high bit rate Intel Confiden?al 18
P.120x.x Video Impairment Model P.1202 Approach Design a metric for evalua?ng quality in specific instance; i.e., IPTV services This significantly constrains the problem, and should make the solu?on much more feasible Known video encoder; encoder output available Known channel; impairment paiern available Four main video distor?ons accounted for Compression Ar?facts Due to lossy encoding Slicing Ar?facts Due to Packet Loss Concealment (PLC) of lost packets Freezing Ar?facts Due to PLC replacing erroneous frames with last good frame ( freezing with skipping ) Rebuffering Ar?facts Due to PLC repea?ng a frame un?l frame recep?on recommences ( freezing without skipping / spinning wheel )
P.1202 (ex. P.NBAMS) Two applica?on areas : P.1202.1,"lower resolu?on mode : Same as P1201.1: i.e., QCIF- QVGA- HVGA, mostly for mobile TV and Streaming P.1202.2, "higher resolu?on mode : Linear broadcast TV & Video on- demand: S?ll under Study Intel Confiden?al 20
P.1202 (ex. P.NBAMS) Packet Headers and bitstream input only Not intended for codec evalua?on Not intended for streams with significant rate adapta?on Video Pearson correla?on of 0.918 for P.1202.1 (982 samples) Validated Test Factors Recommenda?on P.1202.1 P.1202.2 Audio BR NA NA Video BR 0.05 6 Mbps Under study Packet loss ü Under study Re- buffering ü Video Resolu?on HVGA, QVGA, QCIF Under study Video encoding H.264/AVC baseline Under study FR s and key frame rates Frame rate 5-30 Hz GOP lengths 2-10s Under study Protocol UDP- based Under study Protocols not tested* TCP- based Under study * (can be used, but may not be reliable) Intel Confiden?al 21